Do you have these questions?

• As a NER researcher, don't you wonder which problems (bottleneck) hold back the progress of this task?

• Does this excellent performance (with BERT) imply a perfect generalization model, or are there still some limitations?

• Do we have a perfect dataset or perfect evaluation methodology?

• How to quantify the interesting phenonmen in the task of NER?

• What's the next or the right direction for the NER?

We help you to find the answer!

论文地址：http://pfliu.com/InterpretNER/rethink-ner.pdf

Notable Conclusions

• The fine-grained evaluation based on our proposed measure reveals that the performance of existing models (including the state-of-the-art model) heavily influenced by the degree to which test entities have been seen in training set \textit{with the same label}.

• The proposed measure enables us to detect human annotation errors, which cover the actual generalization ability of the existing model. We observe that once these errors are fixed, previous models can achieve new state-of-the-art results, $93.78$ F1-score on CoNLL2003.

• We introduce two measures to characterize the data bias and the cross-dataset generalization experiment shows that the performance of NER systems is influenced not only by whether the test entity has been seen in the training set but also by whether the context of the test entity has been observed.

• Providing more training samples is not a guarantee of better results. A targeted increase in training samples will make it more profitable.

• The relationship between entity categories influences the difficulty of model learning, which leads to some hard test samples that are difficult to solve using common learning methods.

Dataset

• PLONER (Person, Location, Organization NER) is purposed to evaluate the cross-domain generalization. We pick the samples which contain at least one of three entity type (Person, Location, Organization) from representative datasets, such as WNUT16, CoNLL03, OntoNotes 5.0.

• ReCoNLL is revised on CoNLL-2003. We manually fixed errors with the instruction of the measure ECR (entity coverage ratio) proposed by the work. Specifically, we corrected 65 sentences in the test set and 14 sentences in the training set.